It’s the latest buzz in artificial intelligence.
Scientists at the Energy Department’s Argonne National Laboratory have pioneered a cutting-edge neuromorphic computer chip—modeled off the brains of bees, fruit flies and other insects—that can rapidly learn, adapt and use substantially less power than its conventional computer chip counterparts.
The physicist leading an interdisciplinary team that developed the state-of-the-art design recently spoke to Nextgov about the chips’ potential to advance artificial intelligence.
“If we start from a biology standpoint, we use ourselves, humans, as a model for intelligent systems, of course. But there are many other branches that evolution has taken where you can sort of reach big computational power,” Angel Yanguas-Gil, principal materials scientist in Argonne’s Applied Materials division, said. “Insects are one of these areas.”
Limitations in contemporary computing motivated the team to look for alternative scientific approaches to apply to microelectronics and the materials that make them up. Over the past 70 years, humans have gotten much better at making physical computing components and it’s at a point where it’s becoming difficult to continuously improve physical devices, he said. Personal devices are simultaneously getting smaller and smarter, and industry watchers anticipate an impending boom in smart sensors that will require less power to compute. And machine learning and deep neural networks are also evolving, but Yanguas-Gil noted they are generally narrowly trained and limited in flexibility.
But where traditional AI and computing systems fall short, the complexities and viability of insects offer a good model system and a modern sense of promise.
“I tend to see an insect as an autonomous system that consumes milliwatts of power, and that in the case of the bee, they can go miles and come back, they essentially know their way in and their way back, they can solve problems, and they can learn tasks that they haven’t been exposed to, even from an evolutionary standpoint,” Yanguas-Gil said.
Evolution has been very ruthless to insects, and people have come to understand that smart insects starve sooner because they consume more energy, he explained. This essentially allows insects to compute things while consuming “just a teeny tiny” amount of power. Further, insects also integrate many different sensing inputs, have receptors for different types of chemicals, can process wind and orient themselves based on sunlight, and some can sense the magnetic field. “So with few neurons, they are sensor heavy, they are power efficient and they can operate in a very noisy environment, apparently very proficiently,” he said.
The physicist added that when researchers look aspirationally at the human brain, difficulties arise because of so many neurons. A bee, on the other hand, has 1 million neurons, and a fruit fly has 100,000. Hives, ant colonies and other insect communities also frequently operate, and thrive, in decentralized environments, much like the space of future technologies.
“I think the chip takes advantage of the tools that evolution has given the insects and an idea is that insects do have an internal context and in a much more primitive way they also think about the world and their nervous system reflects some of the type of patterns that you see in the real world,” Yanguas-Gil said. “And that’s something we are bringing into our chips. If we want to make things very efficient we have to take advantage of the symmetries that we have in our world.”
Spending “probably too many hours” studying insect neuroscience—an area outside of his educational background—he and his team initially worked to understand the insect brain from an engineering standpoint as a dynamic system. They then worked to extrapolate the necessary components from the bug world to an application domain and translated that further into hardware.
Ultimately, the team aimed to better understand how to use novel and emerging materials to make chips more computationally efficient. Their efforts to design and simulate a new neuromorphic chip led Yanguas-Gil’s to two pivotal breakthroughs. They were able to use filters and weights that impact neural connections in real time, depending on what the system deems important and they created a new nanocomposite material—tungsten aluminum oxide—that can enable the chip to operate “at power levels far below one watt.”
“If you want to make a chip that behaves like an insect, you need materials that are capable of changing the properties and can then maintain those changes,” he said.
The technology has many real-world applications, particularly in extreme environments. For instance, he said Argonne is working with the city and University of Chicago to install smart city sensors that leverage data and monitor things like around air quality, traffic and the climate. Yanguas-Gil said the chips could support the sensors by learning in real time to detect poisonous gas. The chips can also be put to use in extremely hot areas around nuclear reactors, where, like a literal fly on the wall, they may be able to detect and respond to anomalies like no chips have done before. They may also advance and accelerate the future of space exploration. The new insect chip system could offer a sense of robustness in its architecture that would make failures less catastrophically impactful.
“In space applications, that is a good example because the power is very demanding, its something where you really want to compute with fewer amounts of power,” Yanguas-Gil said. “And the range of missions that are being proposed in the future are going to require a much greater degree of autonomy.”
The team’s research is funded through the Defense Advanced Research Projects Agency, Argonne and Energy. The physicist said the multi-stakeholder support has been instrumental in what’s become one of his most intellectually rewarding endeavors to date. Going forward, Yanguas-Gil and the team will continue to enhance the chips. The physicist added that he sincerely hopes companies or other entities soon reach out to leverage the newly-developed technology on their own and perhaps apply it as a pacesetter for the next generation of microelectronics.
“I think for me—and for the lab—that is the metric of success. I tend to think that the reason why I'm in the national lab, is because you want to have that kind of impact, you want to manufacture or to be able to help new ideas being applied in the real world and particularly in these areas where its very arcane and it doesn't really make sense for a company to invest in research in doing this,” Yanguas-Gil said. “So I think it's a win intellectually, but it’s also a win when you see it has an impact beyond the scientific community.”